You can also find links to interesting journals in the fields of Medical Oncology, Haematology and Biomarkers. Patient-derived organoids PDOs are versatile experimental models for cancer research. Our biobank included at least 10 patients who underwent neoadjuvant chemotherapy.
We aimed to apply this resource for N-of-1 co-clinical trials, for the discovery of novel therapeutic agents and for biomarker development.
- Donate to arXiv?
- Classic Western Collection!
- Peaceful Endings : The NOPOSAM Project!
- Predicting cancer outcomes from histology and genomics using convolutional networks | PNAS.
- Windows 8: Deciding on Upgrading?
We report here our application of the developed PDO models for predicting clinical outcome of colorectal cancer patients. Surgically resected colorectal specimens from both normal and cancerous tissue were digested and embedded in appropriate matrices to form PDO models. Subsequently, all successful PDO models were collected in Siriraj live biobank together with their clinical data. Herein, we characterized our derived PDO models by pathological and molecular techniques. Some models were profiled with panel sequencing to assess for the causal driver mutations.
Drug sensitivity of each PDO model to standard-of-care drugs were generated and compared with clinical outcome in each patient. We successfully established PDO models from different colorectal sites, of different pathological grades and staging. We identified heterogeneous morphologies of PDO models which appear to be corresponding well with the pathological and genomic profiles of each specimen origin.
Although the PDO models from different colorectal cancer patients exhibited variable response to 5-fluorouracil 5-Fu , Leucovorin and Oxaliplatin, the response patterns are consistent with the clinical outcome of the individual patients as determined by tumor regression grade or radiologic imaging for neoadjuvant cases.
We are examining the use of drug sensitivity profiling of PDO models for prioritizing anti-cancer treatment choices in a prospective clinical study. The Siriraj live biobank of PDO from colorectal cancer patients can serve as a useful resource for N-of-1 co-clinical trials, and for comparison of drug response among patients from different ethnic backgrounds.
OncologyPRO structure explained. Daily News OncologyPRO, in partnership with medwireNews, brings you updates each day in the field of medical oncology. What's new on OncologyPRO?
You are here
OncologyPRO structure explained OncologyPRO has evolved in July to offer you direct access to our rich content either by tumour site or according to the objectives of your visit. Bioethics, legal and economic Cancer Aetiology, Epidemiology Cancer Immunology and Immunoth Cancer in Special situations Rare Cancers Resources for Patients. Practice Tools ESMO has developed a set of organ-specific tools designed to assist oncologists in their daily practice.
Our analysis was blinded from the fundamental concepts such as cells and nuclei, different tissue compartments and histological grade that were incorporated in the previous studies. Nevertheless, we were able to train an independent risk predictor based on the training cohort, using the raw image data and follow-up information only.
Outside breast cancer, direct outcome prediction using tissue morphology has been successfully applied in colorectal cancer [ 39 ] and glioma [ 40 ]. Even though pathologists do not perform such a direct risk assessment as part of breast cancer diagnostics, we wanted to evaluate the prognostic potential of morphological features detected by pathologists in a small tumour tissue area a TMA core and compare this with the corresponding digital risk score. The analysis indicated that the visual risk score was a significant predictor of outcome, but that the digital risk score yielded a slightly stronger discrimination than the visual risk score C-index 0.
As expected, the visual risk score correlated with known tissue entities mitoses, pleomorphism, tubules, necrosis and TILs. Interestingly, the DRS group associated only with pleomorphism and tubules, indicating that the machine learning algorithm partly has learned known prognostic entities, but partly has extracted features that are not fully explained by known factors.
This was supported by multivariate survival analysis with DRS and visual risk score, which showed increased discrimination C-index 0. One of the main reasons behind the success of deep learning and CNNs has been improved availability of large data sets [ 41 , 42 ]. The best-performing CNNs for object detection and classification are trained with millions of images [ 43 — 45 ].
Contrary to classification based on handcrafted image descriptors and shallow learners, CNN inherently learns hierarchical image features from data, and larger data set usually leads into more powerful models. This ability to learn features directly from the data makes CNNs perform well and why they are easy to use. However, when only limited number of data points is available, direct end-to-end training of a CNN might not lead into any added benefit over handcrafted features and a shallow classifier. Our goal in the design of the computational pipeline for patient outcome classification was to combine the best from the both worlds; the descriptive power of CNNs with the capability of shallow learners to generate robust models from more limited data set.
Post your comment
Generally, this approach is known as transfer learning, which is a popular strategy to achieve a strong performance even with smaller data sets [ 46 , 47 ]. We took advantage of a CNN trained on the ImageNet [ 48 ], a large database for object recognition, and used it for extracting local image descriptors. An important benefit of this approach is less computational requirements, since training of the CNN is not needed.
Furthermore, the approach is agnostic with regard to the CNN used and is easily amendable and compatible with novel model architectures frequently discovered and shared online for the research community. The ImageNet consists of photographs representing natural objects from bicycles to goldfish. IFV is an orderless descriptor aggregation method, capturing the first- and second-order statistics of the GMM modes. The GMM modes were learned in the training set of tumour tissue images, and therefore this intermediate unsupervised learning phase further fine-tunes the features more suitable to the domain of histological images.
Our study has some important limitations. The cohort used in this study was centrally scanned using the same slide scanner and therefore the generalisation of the outcome prediction to tissue images from other slide scanners was not taken into consideration. Moreover, our study considered only small tumour area in the form of a TMA spot image. Although our analysis indicated correlation with the computerised prediction and pleomorphism and tubules, a major limitation of the current work is the difficulty to explain the exact source and location of the predictive signal, i.
Some approaches to answer this shortcoming have been presented [ 50 ], but this is an active research question in field of machine learning and no direct solution for this exists at present. We intend to address this in the future studies. Our findings indicate that computerised methods offer an innovative approach for analysing histological samples. Nevertheless, future studies are required to validate our findings, test similar algorithms on larger data sets representing different malignancies.
We have demonstrated how machine learning analysis of tumour tissue images can be utilised for breast cancer patient prognostication. Our results show that it is possible to learn a risk grouping, providing independent prognostic value complementing the conventional clinicopathological variables, using only digitised tumour tissue images and patient outcome as the endpoint. These findings suggest that machine learning algorithms together with large-scale tumour tissue image series may help approximate the full prognostic potential of tumour morphology.
Development of the methodology: RT. Study supervision: NL, JL. Other authors have no conflict of interest.
Publisher's Note. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. National Center for Biotechnology Information , U. Breast Cancer Research and Treatment. Breast Cancer Res Treat. Published online May Panu E. Author information Article notes Copyright and License information Disclaimer.
Donate to arXiv
Received Dec 10; Accepted May Abstract Purpose Recent advances in machine learning have enabled better understanding of large and complex visual data. Results In univariate survival analysis, the DRS classification resulted in a hazard ratio of 2. Conclusions Our findings demonstrate the feasibility of learning prognostic signals in tumour tissue images without domain knowledge.
Electronic supplementary material The online version of this article Keywords: Breast cancer, Machine learning, Deep learning, Outcome prediction. Background There is a growing interest around the potential of machine learning to improve the accuracy of medical diagnostics [ 1 ].
- The Planning Imperative and Human Behavior.
- Quicklet On Jules Vernes Around the World in Eighty Days.
- Leadership and Global Justice (Jepson Studies in Leadership).
- Поэтические дневники:Тетрадь 3:Тетрадь 3 (Catalan Edition).
- Belem, Portugal: 4 Centuries in 4 Hours (Beds I Have Slept In: A Lifetime of Travels Around the World Book 3)!
- New Guidelines Should Improve Cancer Outcome-Prediction Tools?
Materials and methods Patients and preparation of tumour tissue microarrays We pooled two data sets for the study, the FinProg series and a similar single-centre series from Helsinki University Central Hospital. Outcome classification We extracted local convolutional image descriptors for each TMA spot image by reading the activations from the last convolutional layer of convolutional neural network VGG trained on the ImageNet database [ 25 ], and used improved Fisher vector IFV encoding [ 26 ] to aggregate the descriptors from the image foreground regions into a single descriptor.
Applications of Machine Learning in Cancer Prediction and Prognosis
Visual risk scoring Three pathologists scored the test set TMA spot images into low and high-risk groups using a web-based viewing and annotation software WebMicroscope, Aiforia Technologies Oy, Helsinki, Finland. Statistical analysis The Kaplan—Meier method was used for estimating the survival function [ 32 ] and the log-rank test was used in comparison of survival curves. Results Outcome classification We trained the outcome classifier using a training set of tumour tissue images, and subsequently classified the test set representing breast cancer patients into low and high DRS groups Fig.
Open in a separate window. Outcome classification and survival analysis We investigated the prognostic value of the DRS grouping with univariable and multivariable survival analysis in the test set. High 2. II or III 3. Positive 0. Positive 1. Given 1. Outcome classification and visual risk score Out of the test TMA spot images, were classified by at least one pathologist as not evaluable due to insufficient amount of cancer tissue or partial spot detachment for reliable risk assessment and were therefore left out from the analyses.
Discussion We found that by utilising machine learning algorithms it is possible to extract information relevant for breast cancer patient outcomes from tumour tissue images stained for the basic morphology only. Conclusions We have demonstrated how machine learning analysis of tumour tissue images can be utilised for breast cancer patient prognostication. Electronic supplementary material Below is the link to the electronic supplementary material. Data availability The data that support the findings of this study are available from the University of Helsinki but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available.
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1.
New Guidelines Should Improve Cancer Outcome-Prediction Tools – Consult QD
Obermeyer Z, Emanuel EJ. Predicting the future—big data, machine learning, and clinical medicine. Furthermore, the CIRI-BRCA prediction is superior to the predictive ability of any factor alone, and patients can be stratified from high to low risk based on this approach. Therefore, CIRI is effective in multiple cancer types. Thus far, Drs. But what about CIRI as a predictive biomarker that identifies differential effects dependent on the treatment?
Typically, biomarkers are gene expression, protein levels, or other molecular characteristics that have a correlation with therapeutic response. However, a quantitative risk assessment may also be used to identify biomarkers that are not dependent on molecular characteristics.
The team looked at the FCR chemotherapy regimen a type of chemotherapy using Fludarabine, Cyclophosphamide, and Rituximab that was used in many of the CLL patients in the available datasets.